Recommenders rank items (products, videos, jobs) for each user from behavior and metadata. Common patterns: collaborative filtering, content-based features, and two-stage retrieve-then-rank pipelines.
Approaches
- Collaborative — users who liked A also liked B
- Content-based — similar item attributes or embeddings
- Hybrid — combine signals; business rules on top
Cold start
New users/items lack history—use popularity baselines, onboarding preferences, or content features until data accumulates.
Filter bubble and diversity
Pure engagement optimization can amplify extremes. Product policy may inject diversity, freshness, or seller fairness—metrics beyond click-through rate.
Important interview questions and answers
- Q: Two-stage ranker?
A: Fast retrieval narrows candidates; expensive model reranks top K. - Q: Exploration in recsys?
A: Show some non-obvious items to learn preferences—bandit methods.
Self-check
- Explain collaborative vs content-based.
- What is cold start?
Tip: Optimize beyond CTR—diversity and fairness matter for long-term trust.
Interview prep
- Cold start?
- New users/items lack history—use popularity or content features initially.
- Two-stage ranker?
- Fast retrieval then expensive rerank on top K candidates.